AWS launches G4 GPU instances for machine learning and graphics rendering workloads.
Amazon Web Services Inc. has announced the general availability of the G4 instance family, a set of six virtual machines optimized for machine learning workloads that succeed the G3 series AWS introduced in 2017. The performance improvement is significant, with the new instances running ResNet-50, a popular image recognition model, up to twice as fast as their predecessors.
The G4 instances are powered by Nvidia Corp.'s Tesla T4 graphics card, which contains nearly 3,000 processing cores, including 320 Tensor Cores specifically engineered to accelerate AI model computations. Five of the six virtual machines in the G4 family come with a single T4 card per instance, while the g4dn.12xlarge provides four chips along with 192 gigabits of memory. The Nvidia silicon works alongside Intel Corp. central processing units, which handle general computing tasks to free up GPU processing power for AI software.
Beyond machine learning, the G4 instances are well-suited for graphically-intensive workloads such as video rendering, thanks in part to the T4 chips' multipurpose architecture. In addition to the 320 Tensor Cores, the graphics card features 40 RT Cores that accelerate the generation of light and shadow effects to speed up visual processing. As AWS chief evangelist Jeff Barr noted, "The T4 GPUs are ideal for machine learning inferencing, computer vision, video processing, and real-time speech & natural language processing."
The instances are currently available in eight of AWS' 22 global data center clusters. AWS plans to expand support to additional regions and introduce a seventh instance variant featuring eight T4 graphics cards, 93 CPU processing cores, and 384 gigabits of memory. For companies requiring even greater computing power, AWS offers its P3 instance series, with the largest configuration providing eight Nvidia Tesla V100 data center chips, each containing more than 5,700 processing cores, of which 640 are Tensor Cores.